Parcel-based orchard recognition using a semi-supervised spatial-spectrum enhancement network under small samples

An accurate and rapid recognition is very necessary for the decision-making on precision agriculture in recent years. However, the conventional deep learning approaches are limited to small-sample conditions. Particularly, the high accuracy is often required in the spatially heterogeneous agricultural landscapes. In this study, a generalized intelligent recognition framework was developed for the county-scale and parcel-level orchard mapping. Farmland parcels were initially delineated to select a Boundary-Enhanced Segmentation Network (BsiNet) that was optimized by transfer learning on high-resolution GF-2 imagery. There were some variations in the shape, size, and background complexity after precise identification of the parcel boundaries. The resulting parcels were then spatially aligned with Sentinel-2A multispectral imagery in order to extract the spectral information for each parcel, including multiple visible, near-infrared, and shortwave-infrared bands. A high-quality parcel-level dataset was produced to integrate the precise spatial geometry with rich spectral features, thus providing a solid foundation for the subsequent classification. A lightweight depthwise separable convolutional residual network, termed LiteTransResNet, was designed to enhance the extraction and fusion of deep spatial–spectral semantic features. The network was incorporated with the multi-level spatial–spectral feature enhancement modules, residual connections, and depthwise separable convolution, in order to improve the representational capacity. As such, both local spatial structures and subtle spectral variations were captured to maintain the computational efficiency suitable for the large-scale agricultural applications. A semi-supervised learning was employed to avoid the scarce labeled data. Consequently, the high-confidence predictions from unlabeled parcels were selected to dynamically expand the labeled dataset. As such, the dependency on manual annotation was reduced to significantly enhance the generalization of the model for the cross-regional orchard classification. The framework, combining BsiNet-based parcel delineation with semi-supervised LiteTransResNet classification, was integrated to deploy into the plum orchards in Jiashi County, Xinjiang, China. A field survey was recorded to compare with the planting area statistics in order to validate the framework. The overall classification accuracies of 99.11% and 99.46% were achieved, respectively, in Baren and Mixia town, which substantially outperformed the conventional deep learning, indicating the high effectiveness of spatial–spectral semantic feature enhancement with semi-supervised learning under limited samples. Furthermore, the plum planting areas were estimated over twelve towns of Jiashi County, when training only on plum samples. The errors ranged from 2% to 9%, highlighting its robust generalization with the minimal training data. The framework also generated the high-resolution farmland parcel data at 1-meter spatial resolution and the orchard distribution maps at 10-meter resolution. Both local spatial patterns and multispectral features were effectively captured to facilitate the accurate county-scale orchard mapping. The reliable inputs were provided for precision agriculture. The results indicated that the spatial–spectral semantic feature enhancement with the semi-supervised learning substantially improved the classification performance under small-sample conditions. While the lightweight network structure ensured high efficiency and scalability for the extensive mapping tasks. Overall, a reliable approach was presented for the high-precision, parcel-level orchard recognition and spatial mapping under constrained available samples. The county-wide orchard distribution was accurately estimated using minimal labeled data. Strong support was obtained for the regional-scale orchard monitoring, crop type mapping, growth, and yield estimation in precision agriculture. Additionally, the framework also offered a transferable solution to the high-resolution, small-sample agricultural remote sensing. There were robust generalizations and scalable performance over heterogeneous landscapes. The high-resolution spatial information, multispectral features, and adaptive training were integrated to enhance both the accuracy and operational efficiency of the large-scale orchard classification and monitoring. These findings can also provide a practical and replicable reference for precision agriculture.